Proceedings · Session S-819 · filed November 17, 2026
AI & Emerging Tech in R&DSession paper
Where Pharma Should Hit the Brakes on AI, According to Open Biopharma
Open Biopharma's Richard Jaenisch warns that self-taught AI skills threaten GMP work, and urges pharma to design systems with human review deliberately built in.
By Priya Raman3 min read638 words
Summary
- Richard Jaenisch of Open Biopharma called self-taught AI training without expert guidance 'unequivocally not really a good idea'
- AI models can only train on existing data, while drug development works at the edge of what is known, Jaenisch argued
- He likened proper AI training to a driver's license with specific benchmarks, like SOPs governing GMP work
- Content-generation AI skills do not translate to validation work, and applied-AI job descriptions remain inconsistent industry-wide
Pharmaceutical companies that automate every process risk turning their remaining manual checks into bottlenecks, Richard Jaenisch, senior director of Education, Outreach, and Digital Experience at Open Biopharma, warned in the latest episode of PharmTech's Manufacturing Intelligence podcast. His verdict on self-taught AI skills in regulated work was blunt: using AI to train yourself on AI, when you cannot verify the information and no expert guides you, is "unequivocally not really a good idea."
The episode, hosted by Jaenisch and PharmTech Associate Editorial Director Chris Cole, tackles a question that applies to any AI deployment in a regulated environment: where should life sciences companies slow down and keep human oversight?
What is the 'Luddite Compass'?
Jaenisch framed his answer through what he calls his "Luddite Compass" — a method for judging which forms of work automation should replace and which it should not. History, he noted, offers a long record of technology displacing labor. Scribes gave way to typesetters after the Gutenberg press. Hand drafting gave way to Computer-Aided Design, and CAD itself now yields to tools that convert images directly into CAD designs.
Biology complicates that pattern, he argued. AI models can only be trained on what already exists, while much of drug development works at the edge of what is known. The technology's training data, in other words, does not cover the frontier where much of the industry's actual work happens — a structural limit that portfolio and validation managers should weigh before assigning AI to novel-process decisions.
Why do attitudes matter as much as algorithms?
Human readiness is just as uneven as the technology, according to Jaenisch. Most people occupy an ambivalent middle ground: they distrust AI while still using it for personal tasks. New hire cohorts can contain people who refuse to touch the tools alongside a smaller group that feels a near-compulsion to use them.
That second group concerns him. Heavy reliance can produce employees who cannot work without the tool — a dependency that becomes a staffing and compliance risk when the tool fails, hallucinates, or reaches beyond its validated scope.
His model for proper AI training is a driver's license: a credential demonstrating that specific benchmarks have been met, comparable to the standard operating procedures that govern GMP work. Informal, unguided learning does not meet that bar.
What should hiring managers change?
For hiring managers, the lesson is to interrogate what "AI skills" actually means on a résumé. Someone skilled at producing content with AI tools will not necessarily be useful for validation work, Jaenisch cautioned. The competencies are distinct, and treating them as interchangeable can put the wrong person in a GMP-critical seat.
Job descriptions for applied AI roles remain inconsistent across the industry, he added, which compounds the problem: candidates with the same title on paper may carry very different capabilities in practice.
His practical advice is deliberately low-tech. Plan first and "let your brain do some work" before reaching for the tool. Teams that skip the planning step and let AI run every process discover that their residual manual checkpoints become the constraint on throughput — the opposite of the efficiency the automation was meant to deliver.
How should companies design the system?
The answer, Jaenisch said, is architectural rather than reactive. Companies need to design the whole system with human review deliberately built in from the start, not bolted on after automation has already reshaped the workflow.
The position carries direct implications for R&D and manufacturing budgets: resources must continue to fund trained human reviewers, structured AI credentialing, and consistent role definitions even as vendors pitch end-to-end automation. Jaenisch's closing guidance — plan, verify, and license the humans as carefully as the machines — sets the agenda for regulated AI adoption as the industry's applied-AI hiring wave continues.
via pharmtech.com (Source)
Filed under
- artificial-intelligence
- human-oversight
- pharmaceutical-manufacturing
- gmp-compliance
- ai-training
More from Priya Raman
Show full bio
Staff writer covering money, funding and market shifts around research and development at Hypothesis Wire.
281 articles
References
- Anthropic's AI Lab Sparks Biology Backlash Over Discovery Claim
- Pharma AI Defense Needs Data Stewards, Not New Software
- Anthropic's Claude Moves Into the Lab: AI Now Drives Instruments
- AI Playbook Expected December as Pharma Weighs Vendor Model Change Accountability
- 10x Genomics CTO: AI will leave parts of biomedical R&D 'painfully slow'